New Phi-3.5 Models from Microsoft, including new MoE (huggingface.co)

Mini: https://huggingface.co/microsoft/Phi-3.5-mini-instruct

Large MoE with impressive benchmarks: https://huggingface.co/microsoft/Phi-3.5-MoE-instruct

Vision: https://huggingface.co/microsoft/Phi-3.5-vision-instruct

The Phi models always seem to do really well when it comes to benchmarks but then in real world performance they always fall way behind competing models.

Does anyone have an idea what the output token limit is? I only see mention of the 128k token context window, but I bet the output limit is 4k tokens.